Spatiotemporally Heterogeneous Effects of Urban Landscape Pattern on PM_2.5: Seasonal Mechanisms in Urumqi, China

Abstract

PM_2.5 pollution presents a significant risk to urban habitability. The urban landscape pattern (ULP) serves as a crucial regulator that profoundly influences the spatiotemporal distribution features of PM_2.5. Analysis of the driving mechanisms of the ULP is therefore essential for optimizing urban ecological spatial planning. However, the driving mechanism is dynamic and exhibits seasonal variations. This study selected four landscape metrics and four control variables, developed a geographically and temporally weighted regression (GTWR) model, and examined the spatiotemporal and seasonal effects of ULP on PM_2.5 concentrations in the central urban area of Urumqi (CUA) from 2003 to 2023. The results show the following: (1) Over the past two decades, the four ULP metrics have shown an increasing trend in the CUA. (2) The spatial distribution of PM_2.5 concentrations follows a latitudinal gradient, with higher concentrations observed in the northern regions and lower concentrations in the southern regions, initially increasing and then declining over time. (3) The driving mechanisms of ULP on PM_2.5 exhibited significant variations across different locations and time scales. (4) Seasonal variations arise from pronounced meteorological contrasts and intensified pollution from central heating, which is particularly evident in central CUA.

1. Introduction

PM_2.5 refers to particulate matter with an aerodynamic diameter of less than 2.5 µm, which can remain suspended in the air for extended periods and penetrate deeply into the respiratory tract. Globally, PM_2.5 concentrations are typically heightened in densely populated regions experiencing rapid urbanization and industry [1,2]. Over the past few decades, China’s extensive urbanization and industrial growth have resulted in increased energy consumption and deteriorating air quality [3]. Although regulations on energy conservation and emission reduction have been implemented nationwide since 2013, significant pollution incidents continue to arise throughout winter, particularly in the arid region of northwestern China [4,5,6].

The urban landscape pattern (ULP) reflects the composition and spatial configuration of built structures, which can affect aerosol transport and chemical transformation via multiple pathways—including boundary-layer stability changes, enhanced emissions in high-density areas, microclimate modifications, etc.—thus reshaping urban pollutant distribution [7,8,9]. On the one hand, the high-density built-up area with a large population may intensify traffic emissions and energy consumption [10]. On the other hand, spatial characteristics such as urban connectivity and aggregation can affect pollutant dispersion by modifying microclimate conditions, including urban wind corridors and the urban heat island effect [11]. Although numerous studies have confirmed the PM_2.5-reducing role of green infrastructure [12,13], the mechanisms by which non-vegetated built-up landscape patterns influence PM_2.5 remain contested. Numerous studies have investigated this subject in Chinese cities, but the majority focus on large metropolitan areas and urban agglomerations such as the Yangtze River Delta and the Beijing-Tianjin-Hebei region [14,15]. The findings may not be directly applicable to western inland cities because of the frequent inversion layers, long winter heating periods, and heavily enclosed topography.

A number of built-up area landscape metrics that faithfully capture the spatial features and structural elements of urban morphology are frequently used to evaluate ULP. Frequently used metrics include the aggregation index (AI), largest patch index (LPI), mean patch area (AREA_MN), etc., which signify the extent of spatial aggregation, connectivity, and fragmentation within urban environments [16,17]. Numerous studies have explored the statistical relationship between ULP and air pollutant concentrations using correlation analysis, as well as linear and spatial regression models, to investigate the underlying mechanisms across different spatial scales [18,19,20]. Wang and Ogawa [21] quantified the relationship between urban morphology and PM_2.5 using the Spearman correlation coefficient, suggesting that reducing urban fragmentation, enhancing urban compactness, and maintaining continuous urban forms are conducive to air pollution mitigation. Gao, Yang, and Liang [22] employed spatial exploratory analysis (Moran’s I) and spatial econometric models (SLM/SEM) to investigate the relationship between the landscape pattern of urban construction land and PM_2.5 concentrations. Previous studies have shown that incorporating spatial heterogeneity can improve the explanatory power of regression models and reveal spatial variations in local regression coefficients [23,24]. For example, the geographically weighted regression (GWR) model has been widely used to quantitatively assess the overall impact of urbanization on air quality in China, showing that urbanization intensifies PM_2.5 pollution and that the GWR model outperforms global regression models in terms of model fit and interpretability. Liu et al. [25] integrated spatial regression models with GWR to explore the relationship between urban morphology and PM_2.5, finding that more compact urban construction, lower fragmentation of urban land, and lower density of the road network are conducive factors for improving air quality conditions. Both ULP and PM_2.5 are dynamic and constantly changing, despite the fact that these studies offer insightful information on how ULP influences PM_2.5 concentrations. The complexity is unlikely to be revealed through dynamic analyses that rely solely on a single temporal scale. To better understand the spatiotemporal evolution of urbanization’s impact on air quality, researchers have introduced the time dimension into the GWR model, proposing the geographically and temporally weighted regression (GTWR) model to account for spatial and temporal non-stationarity simultaneously [26]. Zhu et al. [27] applied the GTWR model to analyze the spatiotemporal driving mechanisms of ULP on PM_2.5 across 362 prefecture-level cities in China from 2001 to 2020, revealing significant spatial and temporal heterogeneity in the influence of landscape indices. In comparison to GWR, GTWR demonstrated superior performance in capturing the variations in their effects. While Gou, Li, and Wang [28] further used the GTWR model to investigate the impact and spatiotemporal heterogeneity of green space landscape pattern changes on PM_2.5 concentrations in Chongqing from 1980 to 2020.

Prior research has provided substantial insights into the impact of ULP on PM_2.5 concentrations. However, the long-term dynamics of the driving mechanisms of ULP on PM_2.5 remain limited, and analyses of its seasonal mechanisms are conspicuously lacking. Furthermore, the majority of current research ignores the regulating function of fine-scale urban spatial structures in determining the distribution of pollution inside cities by using entire cities as the fundamental spatial unit [29,30]. This study takes Urumqi, a representative continental arid city in western China, as a case study, focusing on the spatiotemporal driving mechanisms of ULP on PM_2.5 within the central urban area from 2003 to 2023. First, the annual spatial distribution patterns of PM_2.5 were visualized using seasonal raster data. Then, four landscape pattern metrics—number of patches (NP), mean patch area (AREA_MN), aggregation index (AI), and Shannon’s evenness index (SHEI)—were calculated in 2003, 2008, 2013, 2018, and 2023. Combined with four control variables, the relationships between PM_2.5 and ULP were examined at the five time points using ordinary least squares (OLS), the geographically weighted regression (GWR) model, and the geographically and temporally weighted regression (GTWR) model. The results from each model were compared and analyzed. Finally, the GTWR model was employed to reveal the spatiotemporal effects of ULP on PM_2.5 concentrations. The overall workflow chart is illustrated in Figure 1. This study explores the evolving relationship between urbanization-driven landscape changes and air pollution variations in arid urban regions, seeking to understand their spatiotemporal linkages and offer insights for developing sustainable environmental management approaches.

2. Materials and Methods

2.1. Study Area

The central urban area of Urumqi (CUA) has been selected as the study area, which is an arid city in western China. It is located in an inland basin near the northern foothills of the Tianshan Mountains and is characterized by a moderate continental climate, which is susceptible to PM_2.5 pollution. The specific scope is based on the Urumqi Master Plan (2011–2020). The CUA primarily comprises six administrative districts of Urumqi: the New Urban District, Toutunhe District, Tianshan District, Shayibake District, Midong District, and Shuimogou District, which are crucial to urbanization and industrialization. Figure 2 shows the study area’s extent alongside its 2020 land-use/land-cover (LULC) classification. Over the past two decades, Urumqi has experienced swift urbanization and a marked rise in industrial output and residential development, leading to considerable changes in the ULP and severe PM_2.5 air pollution, especially in winter when unfavorable meteorological conditions exacerbate pollution accumulation.

2.2. Data Sources

The PM_2.5 concentration data utilized in this study are derived from the China High Air Pollutants (CHAP), encompassing a long-term series of PM_2.5 concentration raster data for the whole of China spanning from 2000 to 2024, with a spatial resolution of 1 km and a monthly temporal resolution [31,32]. This dataset has been extensively employed in atmospheric and other research fields, with its accuracy rigorously tested. The land-use data are derived from the annual land cover dataset of China Land Cover Dataset (CLCD), which is based on the Landsat data on Google Earth Engine. This CLCD encompasses the yearly land cover information of China from 1985 to 2023, with a spatial resolution of 30 m [33].

Meteorological and topographical factors significantly influence the dispersion and deposition of PM_2.5, so incorporating these factors into the analysis is essential for comprehensively understanding the relationship between ULP and PM_2.5 [34,35]. In this study, precipitation, air temperature, wind speed, and terrain factors were introduced as control variables in the model to minimize the interference of these factors [28,36,37]. The meteorological variables were obtained from the National Tibetan Plateau Data Center, and seasonal values were calculated for each variable. The topographical variable utilizes the digital elevation model (DEM) data, which are derived from the United States Geological Survey (USGS).

This study defines winter as December of the current year through February of the subsequent year and summer as June to August of the current year. Furthermore, both PM_2.5 and meteorological variables were averaged from their monthly values to generate seasonal means for each year.

2.3. Urban Landscape Pattern Metrics

The ULP is quantified by computing landscape pattern metrics that represent the characteristics of the urban landscape comprehensively and is extensively applicable in PM_2.5 pollution studies. Based on the existing literature, we selected built-up land landscape pattern metrics that capture urban structural connectivity, aggregation, and diversity. Meanwhile, all selected variables were retained after collinearity diagnostics (VIF < 10) to ensure non-redundancy. Four metrics were selected in this study, including three class-level metrics for the impervious-surface patches, mean patch area (AREA_MN), aggregation index (AI), and number of patches (NP), and one landscape-level metric, Shannon’s evenness index (SHEI). The specific formulas of each landscape metric can be found in

Table 1. Calculation formula and details of ULP metrics.

Index	Formula	Unit	Description
Mean Patch Area (AREA_MN)	$AREA\_MN={\displaystyle \frac{{\displaystyle {\sum }_{j=1}^n{\alpha }_{\mathit{ij}}}}{n}}\times {10}^{-4}$	km2	aij is the area of patch j of class i; n is the number of impervious-surface landscape patches in the grid.
Number of Patches (NP)	$NP=n$	count	n is the total number of patches in the impervious-surface category.
Aggregation Index (AI)	$AI={\displaystyle \frac{g_{\mathit{ii}}}{\max \_g_{\mathit{ii}}}}\times 100$	%	gii is the number of adjacent edges between patches of class i, and max_gii is the maximum number of adjacent edges between patches of class i under a fully aggregated configuration.
Shannon’s Evenness Index (SHEI)	$SHEI={\displaystyle \frac{-{\displaystyle {\sum }_{i=1}^m\left(p_i\ln p_i\right)}}{\ln m}}$	None	pi is the proportion of landscape class i in the total area, and m is the total number of landscape categories.

. AREA_MN represents the mean patch area of impervious-surface patches within each grid cell, which can be used to measure the connectivity of construction land and reflect the degree of fragmentation to a certain extent. The AI measures the spatial adjacency of impervious-surface patches within each grid cell. NP denotes the total number of impervious-surface patches per grid cell. Increased NP values generally indicate a more dispersed urban spatial arrangement and a heightened degree of fragmentation. SHEI is utilized to characterize the equilibrium of landscape types. Elevated SHEI values indicate a more equitable distribution of diverse land use categories, whereas a diminished value implies a prevailing category.

2.4. Statistical Models

2.4.1. Spatial Autocorrelation Test

In spatial econometrics, the presence of significant spatial autocorrelation in the dependent variable is a fundamental prerequisite for applying spatial analytical methods. Spatial autocorrelation indicates the degree to which similar values of a variable are spatially clustered or dispersed. This characteristic is typically assessed using the Global Moran’s I statistic [38]. The global Moran’s I was calculated in the following equation:

$$I={\displaystyle \frac{n{\displaystyle {\sum }_{i=1}^n{\displaystyle {\sum }_{j=1}^n{w}_{\mathit{ij}}(x_i-\overline{x})(x_j-\overline{x})}}}{({\displaystyle {\sum }_{i=1}^n{(x_i-\overline{x})}^2}){\displaystyle {\sum }_{i=1}^n{\displaystyle {\sum }_{j=1}^n{w}_{\mathit{ij}}}}}}$$

(1)

where I is the global Moran index; x_i and x_j are the PM_2.5 concentration values at locations i and j, respectively; x is the global mean; w_ij is the element of the spatial weight matrix between locations i and j; and n is the total number of sample points.

2.4.2. Co-Linearity Diagnosis

Multicollinearity, resulting from strong correlations among explanatory variables, may bias parameter estimates in regression models [39]. To address this issue, the variance inflation factor (VIF) was calculated as follows:

$$VIF={\displaystyle \frac{1}{(1-c^2)}}$$

(2)

where c is the coefficient of determination, representing the variance of one explanatory variable explained by the others in the model. While VIF < 10, it can be deduced that multicollinearity between the respective variables is not a significant concern.

2.4.3. Global Linear Regression

The ordinary least squares (OLS) model was used to explore the overall linear relationship between the ULP and PM_2.5 concentration with the following equation:

$$y={\mathit{\beta }}_0+{\displaystyle \sum _{k=1}^n{\mathit{\beta }}_k}{x}_k+\mathit{\epsilon }$$

(3)

where y is the dependent variable PM_2.5 concentration, x_k is the independent variable, β₀ is the intercept term, β_k is the regression coefficient, and ε is the random error term.

2.4.4. GTWR Model

In contrast to the OLS model, the GWR model accounts for spatial heterogeneity by allowing regression coefficients to vary across geographic locations. The GWR model’s mathematical equation is expressed as follows:

$$y_i={\mathit{\beta }}_0(u_i,v_i)+{\displaystyle \sum _{k=1}^n{\mathit{\beta }}_k}(u_i,v_i)x_{\mathit{ik}}+{\mathit{\epsilon }}_i$$

(4)

where (u_i,v_i) is the spatial coordinates of sample i, β₀(u_i,v_i) is the local intercept, and β_k(u_i,v_i) is the local regression coefficient that reflects the characteristics of the relationship between the independent and dependent variables at each spatial location. ε_i is the error term. However, given that the complex interplay between the independent and dependent variables is influenced by both spatial and temporal factors, both OLS and GWR models are insufficient for capturing spatiotemporal heterogeneity simultaneously [26]. The GTWR model incorporates a temporal weighting function, and its equation for calculation is

$$y_i={\mathit{\beta }}_0(u_i,v_i,t_i)+{\displaystyle \sum _{k=1}^m{\mathit{\beta }}_k}(u_i,v_i,t_i)x_{\mathit{ik}}+{\mathit{\epsilon }}_i$$

(5)

where (u_i,v_i,t_i) represents the spatiotemporal coordinates of the sample i. Consequently, the local regression coefficient β_k depends not only on the spatial location of the sample (u_i,v_i) but also on the sampling time t_i.

3. Results

3.1. Spatiotemporal Characteristics of PM_2.5

PM_2.5 concentrations in CUA exhibited obvious temporal and seasonal variations from 2003 to 2023 (Figure 3). In summer, the PM_2.5 concentration exhibited an upward trajectory from 2003 to 2012, reaching a peak of 24.6 μg/m³ in 2012. Thereafter, it underwent a substantial decline from 2012 to 2023, reaching a minimum of 9.4 μg/m³ in 2022. This decline was attributed to the ‘Coal to Gas’ policy implemented in Urumqi since the year 2012 [40]. In winter, the PM_2.5 concentrations ranging from 44.2 to 109.6 μg/m³ (approximately 3.2 to 4.6 times higher than those in summer) exhibit an initial increase followed by a decline as well as summer, with pronounced fluctuations between 2010 and 2016. This variability is primarily attributed to the significant influence of low wind speeds, temperature inversions, and residential heating on inner-city air quality [41]. The marked seasonal differences in meteorological conditions directly affect PM_2.5 fluctuations. As is shown in

Table 2. Descriptive statistics of meteorological variables.

Control Variables	Season	Metric	2003	2008	2013	2018	2023
Precipitation (mm)	Winter	max	9.20	4.50	5.60	8.90	13.80
		min	4.70	2.70	3.10	5.30	8.70
		mean	6.72	3.58	4.21	6.86	10.76
	Summer	max	60.10	26.60	35.40	15.60	20.60
		min	33.00	15.50	22.00	9.20	11.30
		mean	42.35	20.17	27.62	11.69	14.99
Temperature (°C)	Winter	max	−9.40	−13.70	−8.80	−13.90	−10.00
		min	−15.00	−20.70	−15.20	−21.20	−16.90
		mean	−12.88	−17.75	−12.52	−18.15	−13.97
	Summer	max	25.10	27.90	26.30	26.80	28.80
		min	20.20	22.70	21.40	22.00	23.90
		mean	22.89	25.59	24.07	24.56	26.54
Wind Speed (m/s)	Winter	mean	5.72	5.92	5.48	5.53	5.37
	Summer	mean	9.16	9.70	9.32	9.55	9.77

, the average wind speed in summer reaches 9.50 m/s, and the average monthly precipitation is 23.37 mm in the CUA, which is conducive to the dilution and clearance of pollutants. Conversely, inversion is prevalent during winter, characterized by an average wind speed of merely 5.60 m/s, an average temperature decreasing to −15.05 °C, and an average monthly precipitation of only 6.43 mm, which restricts pollutant dispersion and results in significant PM_2.5 accumulation near the surface.

The spatial distribution of PM_2.5 (Figure 4) reveals that elevated concentrations are primarily located in the northern and northwestern parts of the CUA, predominantly within Toutunhe District. This area includes vast farmland and various natural landscapes that could produce significant dust. Furthermore, the low topography and consistent air conditions prevent the dispersion of pollutants, hence intensifying the concentration and retention of PM_2.5 [42]. Conversely, PM_2.5 concentrations in southern CUA are relatively lower due to its higher elevation with better ventilation. Moreover, the extensive vegetation may facilitate the decrease in PM_2.5 concentrations [43].

The spatial autocorrelation of PM_2.5 concentrations was evaluated using Global Moran’s I (

Table 3. Global spatial autocorrelation analysis results.

Season	Year	Moran’s I	p-Value	z-Score
Summer	2003	0.916	0.000	69.083
	2008	0.922	0.000	69.594
	2013	0.924	0.000	69.710
	2018	0.948	0.000	71.523
	2023	0.910	0.000	68.640
Winter	2003	0.947	0.000	71.448
	2008	0.976	0.000	73.632
	2013	0.989	0.000	74.612
	2018	0.970	0.000	73.191
	2023	0.957	0.000	72.164

). In summer, Moran’s I value was 0.9158 in 2003, fluctuated slightly thereafter, peaked at 0.948 in 2018, and declined to 0.9099 in 2023. In winter, Moran’s I remained high across all periods, from 0.9472 in 2003 to 0.9568 in 2023, indicating persistent and strong spatial clustering. The trends in Z-values were consistent with those of Moran’s I, further confirming the marked spatiotemporal clustering of PM_2.5 concentrations and providing robust support for subsequent spatial regression analysis. The results showed that PM_2.5 concentrations exhibited significant spatial clustering in all periods, with p-values < 0.001, demonstrating strong spatial autocorrelation.

3.2. Spatiotemporal Characteristics of ULP

All of the ULP metrics were calculated at five temporal intervals: 2003, 2008, 2013, 2018, and 2023, to represent the broader timeframe of 2003–2023. Over the past two decades, all four ULP metrics in the study area have exhibited an overall increasing trend (Figure 5). Among the metrics, AREA_MN has exhibited the most substantial growth, with an increase of 99.06%, suggesting that the impervious surface in the central part of the city has doubled (Figure 5a) and expanded outward from the central CUA (Figure 6a). Over the past two decades, existing small patches have enlarged and merged with adjacent patches in the center CUA, whereas in the peripheral CUA, natural cover has been transformed into impermeable surfaces, resulting in an increase in the mean patch area. The values of NP increased by 50.58% (Figure 5b) and exhibited pronounced spatial heterogeneity; a decrease in NP values in central CUA indicates reduced fragmentation here, while an increase in peripheral CUA suggests an outward trend in urbanization (Figure 6b). Consequently, small patches in the periphery CUA have exceeded those lost to coalescence in the center CUA over the previous 20 years, resulting in a net increase in NP. Temporal changes in the AI demonstrate an increase of 47.26% (Figure 5c), with high-quality areas in the central zone, reflecting a significant increase in urban spatial agglomeration (Figure 6c). The SHEI exhibited the least significant increase at 20.14% (Figure 5d), indicating a modest enhancement in overall landscape diversity. However, the persistent decline in SHEI values within the central CUA suggests that long-term development has led to increased landscape homogenization, reduced fragmentation, and enhanced connectivity (Figure 6d).

3.3. Comparison Between Models

Following the determination of spatial autocorrelation of PM_2.5 within the study area, this research formulated global and spatial regression models to examine the influence of ULP on PM2.5 in the years 2003, 2008, 2013, 2018, and 2023. Subsequent to the VIF and OLS regression significance tests, all of the variables in the models met statistical requirements. To select a more appropriate model, we employed the OLS, GWR, and GTWR models to analyze the summer and winter data, comparing their performance using four metrics: R², Adjusted R², AIC, and CV-RMSE. Where R² reflects the ability of the model to explain the variance of the dependent variable, while Adjusted R² adjusts for the number of independent variables to mitigate bias from excessive variables. AIC quantifies the balance between the goodness-of-fit and the complexity of the model, with lower values signifying superior models. CV-RMSE quantifies the model’s predictive error via cross-validation, where reduced values denote enhanced predictive capability.

Table 4. Comparison of multiple model results.

Model	OLS		GWR		GTWR
Metric	summer	winter	summer	winter	summer	winter
R2	0.257	0.554	0.403	0.784	0.941	0.895
Adjusted R2	0.256	0.553	0.385	0.774	0.940	0.894
AIC	101,429.2	118,371.1	99,008.2	108,663.7	62,913.4	93,514.7
CV-RMSE	16.06	14.15	11.68	6.01	1.72	3.75

demonstrates that the GTWR model surpasses both the OLS and GWR models across all four metrics, corroborating the findings of Ma et al. [44]. Compared to the OLS model, the application of spatial regression models significantly improves model performance, indicating that incorporating local spatial and temporal heterogeneity allows for a more accurate understanding of the driving mechanisms. The GTWR model achieves an R² of 0.941, an adjusted R² of 0.940, an AIC of 62,913.4, and a CV-RMSE of 1.72 in summer. While in winter, the model yields an R² of 0.895, an adjusted R² of 0.894, an AIC of 93,514.7, and a CV-RMSE of 3.75. Therefore, the GTWR model exhibits an augmented capacity to elucidate the correlation between ULP and PM_2.5 in summer. This can be attributed to frequent winter temperature inversions that trap pollutants in the lower atmosphere, whereas robust summer convection and favorable diffusion conditions enable the model to better capture the underlying impact mechanisms. Moreover, winter heating increases fixed pollution sources, affecting the effectiveness of the model’s explanatory variables.

3.4. Response of PM_2.5 to ULP

3.4.1. Response of PM_2.5 to AREA_MN

In summer, the temporal trend of the mean regression coefficients for AREA_MN shows an obvious negative pattern (−0.54–−0.03). Spatially, the local regression coefficients of AREA_MN demonstrate both positive and negative variations and a notable negative correlation between PM_2.5 concentration and AREA_MN is observed in northwestern CUA. The coefficients increased along a northwest-southeast gradient from 2003 to 2013, with a negative-to-positive shift. In 2018, the coefficients shifted to a north-south band pattern, increasing from west to east, also with a negative-to-positive shift (−0.5 → 0 → +0.67). By 2023, the AREA_MN regression coefficients became significantly negative across the study area, with the strength of negative effects increasing from southeast to northwest (Figure 7a,c).

In winter, the effect of AREA_MN on PM2.5 concentration is stronger than in summer temporally, and there is an entire shift to positive in 2018. In terms of spatial distribution, the local regression coefficients of AREA_MN exhibited an overall south-to-north increasing gradient from 2003 to 2013. In 2003 and 2008, a negative-to-positive shift occurred in central CUA, while by 2013, all regression coefficients turned negative, showing a southwest-northeast intensification gradient (−2.05–−3.84). In 2023, the spatial distribution of coefficients remained similar to 2018, while the high-value region shifted southward to the southwestern CUA (Figure 7b,c).

3.4.2. Response of PM_2.5 to NP

The temporal fluctuation indicated that the impact of NP on PM2.5 throughout summer was generally minimal, with more pronounced adverse impacts observed in 2013 (β_k = −0.40). The local regression coefficients of NP demonstrated relative stability from 2003 to 2018, with significant negative effects observed in northwestern CUA and significant positive effects in the northeastern sector. In contrast, the effects in the central and southern sectors were comparatively weak. During 2003–2013, the negative effects gradually intensified, with coefficients spanning −0.44–1.07 in 2003 and −1.95–0.89 in 2013; by 2018, the negative effects began to weaken, with coefficients spanning −0.37–0.37. In 2023, the local regression coefficients of NP underwent a substantial transformation, increasing along a northwest-to-southeast gradient with a distinct negative-to-positive shift (−0.4 → 0 → +0.28) (Figure 8a,c).

Compared to summer, the effect of NP on PM_2.5 in winter is more pronounced. With the exception of 2018, which exhibited overall positive effects (β_k = 0.73), all other years demonstrated negative effects in terms of temporal analysis. During the period 2003–2013, the negative effect of NP shows a spatial feature of decreasing gradient from northwest to southeast. In 2018, the positive effect of NP on PM_2.5 was significantly stronger, with only the northwestern sector displaying a negative effect, and the coefficients formed a north-south band pattern that increased along a west-to-east gradient. By 2023, the effects of NP became negative across the region, exhibiting a radial pattern centered in the urban core and gradually weakening outward (Figure 8b,c).

3.4.3. Response of PM_2.5 to AI

In summer, except for 2023, the PM_2.5 concentration and AI are positively correlated temporally, consistent with findings from Zhang, Zhang, Meng, Wang, Yao, and Li [17]. Spatially, from 2003 to 2013, the coefficients exhibited a north-south band pattern and increased along a west-to-east gradient. In 2018, the positive effects of AI weakened, and the spatial distribution of the coefficients was reversed compared to 2003–2013. By 2023, the coefficients turned negative across the region, with significant negative effects in eastern CUA (β_k < −0.1) and weak negative effects in the northwestern and southwestern CUA (−0.2–0.0) (Figure 9a,c).

In contrast to summer, the effect of AI in winter consistently remained positive temporally (1.19 – 2.11). The local regression coefficients of AI increased along a west-to-east gradient in 2003 and 2008. The spatial distribution of AI regression coefficients reversed in the winter of 2013, with significant negative effects in southeastern CUA and significant positive effects in the northwestern area sector. By 2018 and 2023, the effects of AI become positive across the region, exhibiting a radial pattern centered in the urban core and gradually weakening outward (Figure 9b,c).

3.4.4. Response of PM_2.5 to SHEI

In terms of temporal variation, the effects of PM_2.5 showed a fluctuating trend, with a weak negative correlation in 2003 (β_k = −0.09), followed by consistently positive effects from 2008 to 2023. In terms of spatial distribution, the SHEI had positive effects in northern CUA and negative effects in southern CUA in 2003 and 2008, with a negative-to-positive shift in central CUA. The positive correlation was further enhanced in 2013 (−0.06 – 1.27). Then, the positive effects weakened, and there were significant negative effects in southern CUA (Figure 10a,c).

In winter, the average SHEI regression coefficients are more significant compared to summer. The local regression coefficients of SHEI showed a heterogeneous pattern in winter 2003, with negative effects in parts of the southwestern and northeastern CUA and positive effects dominating the northwestern and southeastern CUA. Notably, by 2008, the spatial pattern underwent a significant change, with the areas exhibiting positive effects expanding. In 2013, the coefficients increased along an east-to-west gradient with an obvious negative-to-positive shift (−0.11 → 0 → +2.95). By 2018, the coefficients exhibited a southwest-to-northeast increasing gradient, with a significant enhancement of positive effects, peaking at 6.21 in northern CUA. In 2023, the intensity of positive effects weakened overall, with significant positive effects in eastern CUA (Figure 10b,c).

4. Discussion

4.1. Analysis of the Spatiotemporal Driving Mechanism of ULP on PM_2.5

Although it has been proved that the effect of ULP on air pollution is divergent across cities, this study found that the variability of the driving mechanism is still evident even within the same city [45]. Considering that the spatiotemporal distribution of PM_2.5 in winter is largely driven by non-structural factors, the effects of UPL on PM_2.5 concentration mainly focus on the summer season. AREA_MN could represent urban connectivity, while NP represents urban fragmentation. Both of the metrics exhibit a significantly negative correlation with PM_2.5 concentrations in northwestern CUA (Figure 7 and Figure 8). This may pertain to extensive agriculture, whereas the rises in NP and AREA_MN suggest a transformation of cropland into impermeable surfaces, hence contributing to a decrease in PM_2.5 concentration. This conclusion is consistent with the research results of Zhu, Tang, Zhou, Li, Liu, Zhang, Zou, Li, and Peng [27] in Northwest China, where harsh ecological conditions and prominent wind-sand issues prevail. As cropland is progressively replaced by built-up areas, surface hardening intensifies, and bare soil exposure declines, effectively reducing dust emissions. Furthermore, urban expansion in these areas has generated significant financial resources and labor, promoting afforestation and wind-sand management initiatives that further reduce PM_2.5 concentrations [46,47]. In northeastern CUA, NP and AREA_MN have positive effects. This may be associated with the presence of the Midong District Chemical Industrial Park, where the increase in NP and AREA_MN may mean the expansion of polluting factories. In addition, this has destroyed the originally connected grasslands and woodlands in the area, weakening their ability to adsorb and settle PM_2.5 and other particulate matter, making PM_2.5 pollution worse. However, the aggravating effects weaken over time. By 2023, the local regression coefficients of NP and AREA_MN in the northeast of CUA have even turned negative. This may be the result of the stringent environmental protection measures and strict regulation of high-pollution industries that have been enforced in Urumqi in recent years. The center of CUA is characterized by a continuous, dense built-up area with optimal connectivity and minimal fragmentation (as indicated by a high AREA_MN value and a low NP value). The regression coefficients of AREA_MN and NP are both negative, suggesting that enhancing urban connectivity and moderately increasing building fragmentation may contribute to a reduction in PM_2.5 pollution. The enhancement of urban connectivity has been proven to optimize the transportation network and functional layout, thereby reducing traffic exhaust emissions [48]. Meanwhile, a moderate increase in the degree of fragmentation in densely built-up areas frequently results in increased green space and vegetation coverage, which contributes to the improvement in local air quality and the reduction in PM_2.5 concentrations [49]. Consequently, the implementation of greening and other sink landscapes should be prioritized in densely populated areas to mitigate the effects of PM_2.5. In the southern CUA, where the altitude is high and the terrain is complex, positive correlations were identified between AREA_MN and PM_2.5 concentration. This correlation can be attributed primarily to the influence of the mountainous terrain characteristics. The smaller AREA_MN resulted in poor connectivity between buildings in the area, thus restricting residents’ travel and reducing traffic emissions [50].

In general, AI has a significant positive effect on PM_2.5 concentration. High-concentration areas frequently denote dense populations, substantial transportation and industrial activities, and increased motor vehicle exhaust and industrial emissions. As a result, these regions demonstrate increased PM_2.5 levels [51]. However, over time, the aggravating effect of AI on PM_2.5 gradually decreased. By the summer of 2023, the coefficients had turned negative overall. This decline may be attributed to the promotion of purification in areas with high concentrations through planning optimization [52]. Additionally, elevated summer temperatures may stimulate atmospheric convection, facilitate diffusion, and mitigate the adverse effects of agglomeration.

SHEI exhibits a substantial positive correlation with PM_2.5, suggesting that an enhancement in landscape homogeneity may result in elevated PM_2.5 concentrations. This phenomenon can be attributed to the increase in factories, roads, and other facilities in northern CUA, which has led to an increase in landscape fragmentation. This fragmentation has resulted in an increase in the emission of pollutants such as PM_2.5 and dispersed distribution, which is prone to causing local pollutant accumulation. Furthermore, the fragmentation of natural landscapes, such as woodlands and grasslands, has been demonstrated to diminish their capacity for dust trapping, thereby contributing to the deterioration of PM_2.5 concentration [53]. The effects of SHEI in southern CUA are significantly negative. This phenomenon may be attributed to the fact that the higher SHEI value reflects the uniform distribution of green spaces, including woodlands, grasslands, and water bodies, within the region. These green spaces have several ecological services and significantly enhance the decrease in PM_2.5 [54].

4.2. Analysis of Seasonal Variations

The PM_2.5 concentrations in Urumqi show distinct seasonal differences between winter and summer (Figure 11). Therefore, it is imperative to analyze the seasonal variations in the correlation between PM_2.5 and ULP. The difference primarily results from (1) the significant contrast between winter and summer meteorological conditions and (2) the amplifying effect of central heating in central CUA.

A comparative analysis of the ULP reveals that, in contrast to summer, winter exhibits a more pronounced effect on PM_2.5 concentration, both in terms of temporal and spatial variations. This can be attributed to the regulatory effect of ULP on air circulation, pollutant diffusion, and deposition, which is more obvious in winter due to the increased severity of air pollution [55]. The impact of central heating on the urban landscape and the PM_2.5 driving mechanism is more prominent in local areas. The center of CUA is characterized by high population density, and there is a notable increase in energy consumption during the winter months, which contributes to the accumulation of PM_2.5 [56]. In the later stages of urbanization, this effect becomes more prominent, resulting in significant differences in the PM_2.5 driving mechanisms in winter and summer. For instance, the effect of AREA_MN was positive in winter and negative in summer in both 2018 and 2023, indicating that the pollution impact of centralized heating in winter outweighs the pollutant mitigation provided by urban connectivity. In addition, the effects of NP and AI in winter both showed a spatial pattern that gradually weakened from the central CUA outward in 2023 (NP had a negative impact and AI had a positive impact), while the coefficient distribution in summer did not demonstrate a comparable tendency. This finding serves to underscore the substantial impact of centralized heating systems in winter on PM_2.5 levels.

5. Conclusions, Limitations, and Future Work

5.1. Major Findings

This study constructed a GTWR model to analyze the relationship between ULP and PM_2.5 concentrations in the central urban area of Urumqi from 2003 to 2023 and further explored the spatiotemporal and seasonal differences in the driving mechanism. The primary conclusions are as follows:

(1)

The ULP has changed significantly from 2003 to 2023 in the central urban area of Urumqi. The trend of urban expansion is obvious, and the four ULP metrics have exhibited an overall upward trajectory, reflecting the continuous enhancement of urban continuity and agglomeration. Concurrently, the landscape diversity surrounding the CUA has undergone an increase.

(2)

The PM_2.5 concentration in Urumqi changed significantly from 2003 to 2023, with initially rising and subsequently declining. This trend was primarily attributed to the implementation of environmental protection policies, such as Urumqi’s ‘Coal to Gas’ initiative initiated in 2012. Moreover, seasonal variations were pronounced, with winter concentrations significantly exceeding those in summer.

(3)

The influence of ULP on PM_2.5 concentrations exhibits significant spatiotemporal variability, even within the same city. The mechanism varies significantly based on geographical location and time. Consequently, the regulation of PM_2.5 pollution should take into account the variations in functional zoning and geographical space within the city.

(4)

The effect of ULP on PM_2.5 has significant seasonal differences, which are mainly due to meteorological conditions and heating periods. The difference caused by meteorology is more obvious overall, while the difference caused by heating is more prominent in central CUA. Consequently, the management of PM_2.5 must prioritize the control of seasonal factors, particularly during the winter months, when PM_2.5 pollution issues are frequently more severe.

5.2. Limitations and Future Work

Although this study has made some progress in revealing seasonally distinct, spatiotemporally heterogeneous effects of urban landscape patterns on PM_2.5, several limitations remain and warrant further research. Firstly, due to the scarcity of continuous ground stations early in our study, we used the validated CHAP satellite product to fill spatial gaps. Although CHAP is widely cited, we acknowledge that it has not yet been validated against ground-based PM_2.5 observations in our study, which may limit a comprehensive assessment of model accuracy. Future studies will incorporate available ground-based PM_2.5 observations to calibrate and validate the model outputs.

Additionally, due to the lack of continuous building height data at the regional scale from 2003 to 2023, we were unable to include three-dimensional metrics in this study. In future work, we plan to leverage recently available TanDEM-X bistatic interferometry and high-resolution optical stereo imagery to reconstruct digital surface models (DSMs) at multiple time points, extract building height information, and thereby more fully reveal the three-dimensional mechanisms by which urban landscape patterns regulate the atmospheric environment.

Moreover, although meteorological and topographic factors were included in our study, non-spatial variables such as industrial emissions, traffic volumes, and energy consumption for heating were not included, despite often driving the temporal variability of PM_2.5. To address this, future work will integrate socioeconomic and emissions data as covariates to disentangle the relative impacts of urban form and human activities on air quality.

Furthermore, the mechanistic interpretations in Section 4.1 rely primarily on GTWR coefficient patterns and theory. However, the strength of some proposed explanations may be limited due to the lack of key data and empirical evidence. For example, in the northwestern CUA, the strong negative correlations of NP and AREA_MN were attributed to farmland conversion to impervious surfaces. However, this requires supplemental data—such as the spatial distribution of underlying surface types and the proportion of farmland-related aerosol emissions—to clarify how landscape shifts drive aerosol sources. In future work, we will supplement key datasets, such as high-resolution land-use data and aerosol optical depth (AOD) data, to strengthen our explanations.

Lastly, we have not quantitatively decomposed the landscape-pollution feedback mechanisms. Future work will integrate WRF-CMAQ or similar models to assess aerosol transport, deposition, and source apportionment, quantifying how landscape evolution impacts aerosol formation and PM_2.5 deposition. This integrated modeling will offer quantifiable assessments to guide urban planning in arid regions.